MERIT reveals the impact of genomic context on sequencing error rate in ultra-deep applications

Rapid progress in high-throughput sequencing (HTS) and the development of novel library preparation methods have improved the sensitivity of detecting mutations in heterogeneous samples, specifically in high-depth (>500×) clinical applications.

M E T H O D O L O G Y A R T I C L E Open Access

MERIT reveals the impact of genomic

context on sequencing error rate in ultra-deep applications

Mohammad Hadigol1and Hossein Khiabanian1,2*

Abstract

Background: Rapid progress in high-throughput sequencing (HTS) and the development of novel library

preparation methods have improved the sensitivity of detecting mutations in heterogeneous samples, specifically in high-depth (>500×) clinical applications However, HTS methods are bounded by their technical and theoretical

limitations and sequencing errors cannot be completely eliminated Comprehensive quantification of the background noise can highlight both the efficiency and the limitations of any HTS methodology, and help differentiate true mutations at low abundance from artifacts

Results: We introduce MERIT (Mutation Error Rate Inference Toolkit), designed for in-depth quantification of

erroneous substitutions and small insertions and deletions MERIT incorporates an all-inclusive variant caller and considers genomic context, including the nucleotides immediately at 5and 3, thereby establishing error rates for 96 possible substitutions as well as four single-base and 16 double-base indels We applied MERIT to ultra-deep

sequencing data (1,300,000×) obtained from the amplification of multiple clinically relevant loci, and showed a significant relationship between error rates and genomic contexts In addition to observing significant difference between transversion and transition rates, we identified variations of more than 100-fold within each error type at high sequencing depths For instance, T>G transversions in trinucleotide GTCs occurred 133.5± 65.9 more often than

those in ATAs Similarly, C>T transitions in GCGs were observed at 73.8± 10.5 higher rate than those in TCTs We also

devised an in silico approach to determine the optimal sequencing depth, where errors occur at rates similar to those

of expected true mutations Our analyses showed that increasing sequencing depth might improve sensitivity for detecting some mutations based on their genomic context For example, T>G rate of error in GTCs did not change

when sequenced beyond 10,000×; in contrast, T>G rate in TTAs consistently improved even at above 500,000×.

Conclusions: Our results demonstrate significant variation in nucleotide misincorporation rates, and suggest that

genomic context should be considered for comprehensive profiling of specimen-specific and sequencing artifacts in high-depth assays This data provide strong evidence against assigning a single allele frequency threshold to call mutations, for it can result in substantial false positive as well as false negative variants, with important clinical

consequences

Keywords: Deep sequencing, Sequencing noise, Genomic context, Polymerase fidelity, Optimal depth

*Correspondence: h.khiabanian@rutgers.edu

1 Center for Systems and Computational Biology, Rutgers Cancer Institute of

New Jersey, Rutgers University, New Brunswick, NJ, USA

2 Department of Pathology and Laboratory Medicine, Rutgers Robert Wood

Johnson Medical School, Rutgers University, New Brunswick, NJ, USA

© The Author(s) 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0

International License ( http://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made The Creative Commons Public Domain Dedication waiver

The rising utilization of high-throughput sequencing

(HTS) in clinical oncology has transformed our

under-standing of cancer evolution and has provided

clini-cians with an invaluable tool for precise diagnosis and

prognosis

In clinical cancer genomic testing, target-capture library

preparation assays are favored over whole genome or

whole exome sequencing approaches because of their

lower cost in obtaining higher sequencing depth – the

number of reads covering a specific locus [1]

High-depth DNA sequencing enables confident detection of

small clones of somatically mutated cells in

heteroge-nous tumor samples, where in addition to

genomi-cally diverse cancer cells, contaminating normal cells

may also be present Using polymerase chain reaction

(PCR)-based amplicon or hybridization-capture

enrich-ment techniques, clinical-grade cancer sequencing panels

are capable of producing 500 to> 10,000 reads mapping

to each targeted locus [2, 3] Specifically, a minimum

average depth of 500× is strongly advised by regulatory

bodies for reliable detection of somatic mutations with

variant allele frequencies (VAFs) as low as 5% in tumor

specimens [4]

The power to detect small clones in heterogenous

sam-ples may improve by increasing depth; however, confident

detection and differentiation of true mutations with low

VAFs, e.g.,< 0.1%, from the sequencing artifacts remains

a challenge HTS errors are dominated by misreading a

base within the instrument or nucleotide

misincorpora-tions during library enrichment with PCR Differential

rate of substitution errors in HTS has been observed and

attributed to common DNA damaging events such as

spontaneous deamination, presence of oxidized bases in

cells in addition to ex vivo oxidation during DNA

extrac-tion [5], or short-lived high temperatures during acoustic

shearing [6] Such events often lead to higher rates of

tran-sitions versus transversions [7–11] or increased number

of errors in specific genomic contexts These differences

can be more pronounced at higher sequencing depths and

directly impact the sensitivity for detecting true mutations

with low VAFs Here, we hypothesize that the genomic

context of substitution errors, i.e., the nucleotides

imme-diately at their 5 and 3, is a determinant factor in

esti-mating their rates at high sequencing depths To this end,

we generated ultra-deep sequencing data (1,300,000×)

and developed MERIT (Mutation Error Rate Inference

Toolkit), a comprehensive pipeline designed for in-depth

quantification of erroneous HTS calls Using MERIT, we

show a significant relationship between substitution error

rates and their sequence contexts In addition to observing

more than three orders of magnitude difference between

transition and transversion error rates, we identify

vari-ations of more than 130-fold within each error type at

high sequencing depths We also propose an in silico

depth reduction approach to provide insights on estimat-ing optimal depth – where sequencestimat-ing errors exist at rates similar to those of true mutations Finally, we propose an assay for detailed assessment of nucleotide-incorporation fidelity for four high-fidelity DNA polymerase molecules

Methods DNA sample

We obtained HapMap NA19240 human genomic DNA (5μg) from Coriell, purified from immortalized

lympho-cytes using the Qiagen Autopure LS instrument in TE buffer (10 mM Tris, pH 8.0/1 mM EDTA) with con-centration of 301 ng/L We assessed sample quality and concentration using Nanodrop and Qubit dsDNA assays before library preparation

DNA polymerase enzymes and primer design

We used four high-fidelity DNA polymerase enzymes – NEBNext® High-Fidelity 2X PCR Master Mix (Hi-Fi 2X), NEBNext® UltraTM II Q5® Master Mix (Ultra II), KAPA HiFi PCR kits with ReadyMix (KAPA), and InvitrogenTM PlatinumTM SuperFiTM DNA polymerase (SuperFi) – for PCR amplification We designed the primers using Primer3 [12] to target four loci in the TP53 and SF3B1

genes such that the paired-end reads (R1 and R2) are significantly overlapped (Additional file 1: Tables S1 and S2)

PCR amplification, indexing, and sequencing

We performed twenty PCR cycles using the Hi-Fi 2X, KAPA, and SuperFi polymerases, and 16 cycles using the Ultra II polymerase in the first round of amplification (Additional file 1: Table S3) The cycle numbers were determined after initial PCR amplification tests in order

to obtain similar amount of DNA for each enzyme The second round of PCR for multiplexing and cluster gen-eration included seven cycles for all four polymerases (Additional file 1: Table S4) After each PCR amplifica-tion, AMPure Bead cleanup was performed First, 0.4× ratio (20 μL AMPure bead to 50 μL PCR product) was

used to remove gDNA and larger fragments (i.e., > 600

bp) For the saved supernatant, additional 80μL AMPure

Bead was added to bring the total to a 2× ratio The beads were eluted with 22 μL EB (10 mM Tris, pH

8.0) The annealing temperature of 66°C was determined based on the product specificity and yield for all poly-merases after performing gradient PCR optimization at eight different temperatures (Additional file1: Figure S3) Qubit quantification and Bioanalyzer analysis were per-formed for quality assessment Custom amplicon-based sequencing and library preparation were performed at GeneWiz (South Plainfield, NJ) using Illumina HiSeq2500 Rapid Run

MERIT: a comprehensive error rate estimator

Comparative performance analysis of the commonly used

HTS variant callers [13–15] suggests a significant

dis-agreement between their identified variants [16–18]

These differences are mainly rooted in each pipeline’s

specific filtering and statistical methodology For

exam-ple, a number of filters is automatically applied to reads

by HaplotypeCaller implemented in the Genome

Analy-sis Toolkit (GATK) [13] to exclude uninformative reads

from the analysis This practice is aligned with the goal

of the majority of variant callers, which is distinguishing

true mutations from the artifacts However, for a precise

quantification of the sequencing noise, i.e., error rate

pro-filing, all the reads need to be included in the analysis as

the ultimate goal is understanding the nature of artifacts

SAMtools [14] is the basis of a number of alignment-based

variant callers [19,20], and has high flexibility for changes

in its filters Therefore, MERIT uses SAMtools to

iden-tify all positions with alternate alleles from the aligned,

indexed sequencing reads By extracting allele frequencies

of substitutions directly from the Pileup file generated by

SAMtools mpileup, we make sure all reads are included in

the analysis

As MERIT is designed for ultra-deep HTS

applica-tions, the input options of its SAMtools mpileup are

set to accommodate high depths while providing the

users the ability to modify these parameters based on

each sequencing data’s characteristics Additional file 1:

Table S5 summarizes the default input parameters of

SAMtools mpileup versus those used in MERIT These

parameters allow MERIT to probe SAMtools Pileup data

and extract sequencing information for all substitutions,

even when they are present in only a single read amongst

tens of thousands Accurate identification of indels is a

challenging problem [21,22] and beyond the scope of this

work Specifically, SAMtools’s filtering criteria in

intro-ducing and extending gaps, could affect calling complex

indels, especially insertions, rendering error rate estimates

sequencing depth-dependent

Next, MERIT obtains the Phred quality score of base

substitutions as well as the average Phred quality of bases

before and after indels These quantities are not provided

in the VCF files generated by SAMtools Of note, we

observed that the alternate allele and total depths at indel

loci are only accurate in SAMtools’s Pileup files and not

in its VCF Therefore, to ensure allele frequency

accu-racy for both indels and substitutions identified, MERIT

extracts the reference and alternate alleles’ depths as well

as the total depths for all the variants from the Pileup

file MERIT also extracts the position-in-read for all

variants Such information, especially in hybrid-capture

sequencing, helps to better quantify the source of errors

in HTS platforms An optional annotation step is also

available Finally, MERIT obtains the genomic context of

the variants from the reference genome, including the nucleotides immediately at their 5and 3, and estimates error rates for 96 possible single nucleotide substitutions

as well as four single-base and 16 double-base inser-tions/deletions (indels) Details of MERIT’s workflow are shown in Fig.1

Error rate estimation

We used a single HapMap sample to generate ultra-deep data, and although there may be small, uncharacterized variations within initial cell population, we assumed that all detected variants were errors accumulated in library preparation or during sequencing

We considered context-specific erroneous base calls at each locus to follow a binomial distribution More

pre-cisely, the probability of a single nucleotide X i with the

genomic context ZX i Z, Z, Z ∈ {A, C, T, G}, in a specific locus i being misread as Y i , i.e., P ZX i Z→ZY i Z followed

P ZX i Z→ZY i Z= P(x i |n i , p ) =



n i

x i



p x i (1 − p) n i −x i,

Fig 1 MERIT’s workflow MERIT is designed for comprehensive

characterization of the sequencing error rate in ultra-deep HTS applications

where p is the combined PCR and sequencing error rate

and n i and x i are the total read depth and the number

of erroneous calls at position i, respectively Assuming a

position-independent p, the probability of observing m

instances of ZXZ→ ZYZerror within each sample was

then given by

P ZXZ→ZYZ = P

m



i=1

x i|

m



i=1

n i , p



=

m

i=1n i

m

i=1x i



pm i=1x i (1 − p)m i=1(n i −x i ).

(1) (See Remark 1 in Additional file1on the sum of

bino-mial random variables.) For the case of indels, a binobino-mial

model was used to describe the error rate as well, but

instead of categorizing them based on their context, indels

were classified based on the type of inserted/deleted base,

as no differential error rates were observed for

context-specific indels

Polymerase fidelity estimation

The estimated error rate in Eq (1) has a unit of

[error/base] It is also common to report the fidelity of

polymerase enzymes as [error/base/doubling] in the

liter-ature where template doubling d is given by

2d= final DNA amount after PCR

starting DNA amount for PCR.

Since precise amounts of input and output DNA were

known for our experiment in its second round of PCR, we

calculated template doubling and estimated polymerase

replication efficiency as the ratio of template doubling d

over the number of PCR cycles performed (Additional

file1: Table S4) To obtain the total amount of template

doubling after performing two rounds of PCR

amplifica-tion, the total number of PCR cycles were multiplied by

the polymerase efficiency which resulted in 20.83, 16.19,

16.87, and 20.98 total template doubling for the Hi-Fi 2X,

Ultra II, KAPA, and SuperFi polymerases, respectively

Alignment and merging

We cleaned the paired-end (PE) reads of adapters using

bcl2fastq Conversion Software (v2.17), and aligned them

to the reference human genome hg19 assembly using the

Burrows-Wheeler Aligner (BWA) tool [23] (bwa sampe

for PE and bwa samse for merged reads along with bwa

aln) We then merged the PE reads that properly mapped

to the targeted loci In our merging scheme, if R1 and

R2 reads did not match at a base, an N was assigned

for that position We discarded read pairs with smaller

than 50 base overlaps or with more than five mismatches

We calculated Phred quality score (Q) of a successfully

merged locus as the sum of the qualities in R1 and R2

reads since these are independent events; Q is given by

Q = −10 log10p where p is the probability that the base

is called incorrectly Merged reads were then mapped to the reference human genome hg19 assembly, and were fil-tered so that they were uniquely mapped (BWA tags X0:1 and X1:0) Finally, in order to make a fair comparison between the error rate of merged and PE reads, we only considered PE reads that were merged successfully and uniquely mapped Additional file 1: Table S2 represents the average depth of merged and PE reads in different loci

To assess the effect of alternate alignment approaches, we tested Bowtie [24] in addition to BWA to map the merged reads to the reference genome

In silico depth reduction

The sequencing assay was designed to obtain an average depth of> 1, 000, 000× bp, but for some amplicons the

average depth was substantially larger (Additional file1:

Table S2) Therefore, an in silico depth reduction

pro-cedure was performed to reduce the high depths and more importantly, generate enough independent samples

to estimate low error rates confidently It should be noted that one of the main hurdles in error rate estimation of high fidelity polymerases via HTS is the lack of signal as errors occur infrequently with increased fidelity, hence,

a large number of samples is required to accurately esti-mate errors As performing ultra-deep sequencing on a large number of samples is not cost-effective, alternatively,

in silicodata at lower depths can be generated from one ultra-deep sequencing run by randomly selecting reads from the original raw sequencing data

Clinical samples

We obtained 29 hematopoietic samples collected from 9 patients with chronic lymphocytic leukemia, previously analyzed by amplicon deep-sequencing (NCBI BioProject PRJNA411889) These samples were sequenced using a custom 88-gene panel, targeting 92 amplicons on Illu-mina HiSeq (2x150bp) at GeneWiz (South Plainfield, NJ) (Supplementary Table 5 in [25]) The reads were cleaned, merged, and aligned to the reference genome as previously described [25] We removed previously detected germline and somatic mutations to ensure that the remaining vari-ants represented only the errors

Results and discussions Impact of merging reads on context-specific error correction

Independent analysis of R1 and R2 reads at 1,300,000× indicated significant variations in estimated error rates across 96 possible sequence contexts (Fig.2) High error rates and low Phred quality scores observed in R2 rel-ative to R1 may be associated with sequencing errors caused by misreading a base, attributed to image analysis

b

c

d

Fig 2 Estimated context-specific substitution error rates for polymerase Hi-Fi 2X a) R1 reads b) R2 read c), the ratio of error rate in R1 over R2.

P-values were computed by performing a two-tailed z-test d) , the difference between their corresponding Phred quality scores We reduced the

depth of paired-end reads to approximately 1,300,000× through an in silico depth reduction procedure Results were obtained by averaging over

100 independent samples to establish error bars which indicate one standard deviation from the average

biases [26] or phasing/pre-phasing [11] These

sequenc-ing errors that dominated the R1 and R2 profiles can

be distinguished from polymerase errors by merging the

overlapped paired-end reads [27–29] Merging, however,

cannot eliminate errors randomly accumulated during

the amplification processes and present in both reads In

contrast to higher rates of transversion versus transition

errors in paired-end reads (Fig.3a), the remaining

PCR-related errors in the merged reads were dominated by

transitions, often with high Phred quality scores (Fig.3b)

MERIT provides further insight for profiling these errors,

which are the main hurdle in distinguishing real mutations

from sequencing noise:

• Merging R1 and R2 reads lowered all the

context-specific error rates The highest reduction in

rate was observed for GTA>GGA transversions

(5,025± 2, 794×) while GCG>GTG transition errors

only improved by a factor of 1.22± 0.07× Moreover, these improvements were context-specific For example, T>A transversion in GTA trinucleotides

showed substantial reduction (568± 249×) compared

to those in CTAs (1.43± 0.31×)

• Transition errors occurred at higher rates relative to transversions, in agreement with previous reports [7–11] This difference was pronounced further when errors were classified based on their context,

denoting a rate of 1.29± 0.04 × 10−3[error/base] for

GCG>GTG versus that of 2.17± 0.92 × 10−6

[error/base] for GTA>GAA (Fig.3b) MERIT also revealed considerable variation within each substitution type For example, T>G transversions in

GTCs occurred 133.5± 65.9× more often than those

in ATAs Similarly, C>T transitions in GCGs were

b

Fig 3 Merging of overlapped PE reads reduces context-specific error rates a) R1 and R2 reads b) Merged reads Depth of merged reads for

polymerase Hi-Fi 2X were reduced in silico to approximately 650,000× Error bars indicate one standard deviation from the average of 100

independent sub-samples

observed at 73.8± 10.5× higher rate than those in

TCTs (Fig.3b)

• The rate of C>A errors in ACCs was the highest of

all such transversions These errors are linked to the

conversion of guanine to 8-oxoG resulting in

mismatched pairing with adenine [30,31] Oxidation

of guanine to 8-oxoG happens naturally in living cells

and can be increased by DNA damaging factors such

as acoustic shearing [32]

• Merging R1 and R2 can correct for the low quality

erroneous bases associated with sequencing errors

Our analysis suggests that such sequencing errors can

be identified and eliminated based on their quality,

when merging the reads is not possible (e.g., in

hybrid-capture-based sequencing where read pairs

are not designed to necessarily overlap)

Finally, we tested whether an alternative alignment

method, such as Bowtie [24], would affect error rate

estimations, and found minimal changes across the 96

genomic contexts (Additional file1: Figure S4)

Effect of mutation context on amino acid variations

In a single codon, the context-specific rate of error for

each base change directly affects the sensitivity of

detect-ing the resultdetect-ing amino acid variation Our data indicated

that the most commonly mutated residues in TP53 and

SF3B1were often more prone to errors and hence

com-paratively less likely to be distinguished from sequencing

errors For example, in TP53, R248Q and R248W are

among the most common mutations found in cancer

patients [33] The transition base changes that result in

these mutations could be confounded by the HTS errors

at an 8-fold higher rate than the transversion alterations that lead to R248L, and 55-fold higher than those that lead to R248G (Fig 4a) Similarly, the K700E mutation

in SF3B1 is the most frequently mutated residue in the

gene’s exon 10 [34, 35]; it results from a T>C mutation

in a TTC trinucleotide that showed the highest rate of error for a non-synonymous amino acid change in its codon (4.74± 0.42 × 10−5 [error/base]) In contrast, the

comparatively rarer I704F mutation – a T>A in a ATG

ref-erence trinucleotide – had one of the lowest rates of error

in its respective codon (5.15± 1.13 × 10−6 [error/base];

Fig.4b) K700E’s 9-fold higher rate of error than that of I704F indicated marked reduction in its relative detection sensitivity

Optimal sequencing depth

Insufficient sequencing depth reduces the sensitivity of detecting variants and leads to loss of statistical sig-nificance for a confident variant calling [36] Conse-quently, sequencing at higher depths is expected to provide robust error rate estimates and improved sensi-tivities in detecting true mutations Accurate estimation

of optimal sequencing depth, beyond which the inferred background error is not further reduced, not only pro-vides a precise view of intrinsic limitations in HTS assays, but also leads to preserving time and resources by avoid-ing unproductive ultra-deep sequencavoid-ing experiments

To provide insight on optimal sequencing depth,

we performed in silico experiments and estimated

context-specific error rates as a function of depth We randomly selected merged reads and constructed simu-lated sequencing data at depths ranging from 1,000× to 700,000×, with 500 independent replicates at each depth

b

Fig 4 Significant variation in error rates for possible amino acid changes at individual codons a) Six frequently mutated residues in the TP53 gene b) Two hotspot residues in the SF3B1 gene The higher the rate of error for a specific base change, the lower the power to distinguish true mutations

from sequencing artifacts at its position Here, the error rates represent the amplification by the Hi-Fi 2X polymerase Error bars represent one standard deviation from the mean of 100 independent sub-samples

to establish confidence intervals (Fig.5) MERIT showed

that the type of substitution error was an important

deter-minant in estimating the optimal depth (Fig 5a) The

error rate estimates for all transitions as well as C>A

transversions did not significantly change as sequencing

depth increased beyond 200,000×; however, the inferred

rates for the remaining transversions marginally improved

at higher depths More importantly, this analysis

high-lighted the importance of context-specific error profiling

in determining detection sensitivity thresholds for true

mutations For example, at 5000×, the corresponding

error rates for all T>A errors, T>A errors in CTAs, and

T>A errors in GTTs were 2.19± 0.37 × 10−4[error/base],

4.27± 2.28 × 10−4 [error/base], and 1.96± 0.02 × 10−4

[error/base], while at 700,000×, these rates were reduced

to 2.02± 0.73 × 10−5 [error/base], 2.5± 0.99 × 10−4

[error/base], and 2.1± 0.89 × 10−6 [error/base],

respec-tively Selecting a frequency threshold for these variants

at 5000× based on the general T>A rate may not yield significant number of false calls independent of their sequence contexts; however, at depths > 5000×,

set-ting a threshold based on all T>A errors would lead

to substantial false positive CTA>CAA and false

nega-tive GTT>GAT calls, as their corresponding error rates

a b c

Fig 5 Context-specific optimal sequencing depth Substitution error rates are classified based on their type (column a) and context (columns b and c) at nine different depths: 1,000×, 5,000×, 10,000×, 25,0000×, 50,000×, 100,000×, 200,000×, 400,000×, and 700,000× In silico depth reduction

procedure was performed on merged reads, amplified by polymerase Ultra II to an average depth of 1,930,473 × The shaded areas are uncertainty bounds of one standard deviation around the average, derived from 500 independent sub-samples

diverge at high depths, reaching a difference of two orders

of magnitude at 700,000×

It should be noted, however, that SAMtools might

not be able to detect all indels at all depths [21, 22]

Although comparing indel error rates might be only

statistically meaningful at fixed sequencing depths, we

did observe a reduction in estimated rate of error

for single-nucleotide deletions relative to sequencing

depth (Additional file 1: Figure S6) Calling all

com-plex indels, especially when they are present in only

a few reads, may require more sophisticated variant

callers whose results can be combined with

substi-tution calls to obtain a comprehensive error profile

by MERIT

DNA polymerase fidelity estimation

High-fidelity DNA polymerases – equipped with proof-reading – result in fewer base misincorporations in PCR enrichment step, and thus, can reduce HTS error rates The Hi-Fi 2X, Ultra II, KAPA, and SuperFi enzymes are marketed as high-fidelity polymerases, specifically designed for efficient amplification of complex templates such as those with GC-rich regions Their providers have reported a fidelity 100× better than wild-type Taq DNA polymerase [37–40]

We applied MERIT to merged reads at equal depths

of 650,000×, ensuring that the estimated fidelities were not affected by sequencing depth When all errors were included in the analysis, global error rates suggested

that these polymerases performed fairly similarly to each

other, with the highest and lowest error rates belonging

to KAPA and SuperFi enzymes, respectively Specifically,

the global substitution error rates for Hi-Fi 2X, Ultra II,

KAPA, and SuperFi were estimated at 2.66± 0.21 × 10−6,

1.91± 0.19 × 10−6, 6.95± 0.54 × 10−6, and 1.76± 0.25 ×

10−6[error/base/doubling], respectively (Additional file1:

Figure S1a)

Because different assays, quantification methods, and

descriptive units [21,41] are often used to estimate the

polymerase fidelity, comparing the reported rates in the

literature is a challenging task and beyond the scope

of this work More importantly, error rate profiles in

HTS data are reported to be platform as well as batch

dependent [42] For example, using single cell sequencing

technique error rates of 5.3× 10−7 [sub/base/doubling]

and 1.6× 10−5 [sub/base/doubling] are reported in [21]

for Ultra II and KAPA polymerases, respectively In

another study [43], a barcoding sequencing approach

yielded a rate of 4× 10−6 [substitutions/base] for Ultra

II while 2.8× 10−7 [substitutions/base] is reported for

KAPA enzyme in [39] Here, we use MERIT to emphasize

on the importance of context-specific polymerase fidelity

estimation and provide a robust comparison of these

com-monly used high-fidelity enzymes performed on a single

sequencing platform

Relying solely on global error rates for comparing the

replication accuracy of these high-fidelity enzymes may

be misleading [44] Previous HTS-based analyses of

poly-merase fidelity estimation have classified substitutions

into transition and transversion types and have showed

preferential rates of error [21, 41, 44, 45] Additional

file 1: Figure S1b represents such classification of the

substitution errors in our ultra-deep data, providing a

more detailed understanding of the replication fidelity

of these enzymes For example, the global substitution

fidelity of SuperFi was found 3.95± 0.65× better than

that of KAPA’s; however, specific substitution fidelity

dif-fered widely C>G errors of SuperFi were 6.88± 2.16× less

frequent than those of KAPA In contrast, for C>A

sub-stitutions, SuperFi’s advantage over KAPA was reduced to

only 1.85± 0.30×

For a more comprehensive analysis, we used MERIT to

estimate 96 context specific substitutions and observed

substantial variations (Fig.6a) For example, TTA>TGA

error rate of SuperFi was found 132± 35× lower than

KAPA, while for GCG>GAG errors, KAPA performed

just slightly better than SuperFi Such classification of

sub-stitution errors based on their genomic context enabled

us to perform robust statistical comparisons between

the replication accuracy of different DNA polymerases

using Spearman’s rank correlation coefficients presented

in Fig.6d, rather than just comparing them using a single

global error rate Moreover, using the data from multiple

regions of the TP53 and SF3B1 genes, we found

lim-ited change in overall error profiles as the similarities between the genomic content of the amplified amplicons decreased (Fig.7)

Application of MERIT to clinical samples

Sample preservation and library preparation of clinical samples can lead to specimen-specific errors MERIT provides a tool to assess and compare such error profiles Additional file 1: Figure S5 represents the substitution error rates estimated for hematopoietic samples collected from leukemia patients presented in [25] The error rate estimates for these clinical samples showed a high rate

of transition errors similar to previous results from a cell line A major difference, however, was that the C>A

errors proceeded by C and T bases were more frequent than those proceeded by A and G Our data did not show

a preferred 3base trailing the misread C base This high rate of C>A errors has been observed in previous studies

[32, 46], specifically an abnormally high rate of CCG>CAG errors in both tumor and normal samples

from cancer patients [32]

Conclusions

Novel library preparation methods have succeeded in reducing the background sequencing noise, which has led

to improving the sensitivity of detecting true mutations

in heterogenous samples PCR-free library preparation methods [47, 48] have forgone the bias associated with the polymerase base incorporation [49,50], however, the large amount of input DNA required in these techniques

is the main burden for their application in clinical cancer genomic testing As the exponential PCR amplification is

a crucial step in HTS, other techniques have focused on minimizing polymerase errors rather than abolishing the PCR step entirely, including Safe-Seq [6], Duplex-Seq [51], Circle-Seq [52], Cypher-Seq [53], and maximum-depth sequencing [54] Despite all improvements, the back-ground noise is not completely eliminated The additional cost and complexity of these methods as well as their lower yield [54] limit their utilization in clinical cancer genomic testing Specifically, a limited starting material,

as is usually the case for tumor specimens, could results in poor sample representation due to inefficiencies in adapte

r ligation and loss of genetically diverse small clones [55]

In this paper, we provided a comprehensive method for profiling sequencing artifacts and discussed their impact

on accurate variant detection in amplicon-based HTS data We proposed an approach for determining the optimal sequencing depth, where errors occur at rates similar to those of true mutations Our data obtained from Illumina platforms confirmed previous results on the differential rates of errors in paired-end sequencing reads [11], and indicated that merging the overlapping

b

c

d

Fig 6 In-depth comparison of the error rates for four high-fidelity polymerases a) Context-specific substitutions b) Single-base insertions c)

Single-base deletions d) Spearman’s rank correlation coefficient between context-specific error profiles Results are obtained by averaging over 100

independent samples to establish error bars, which indicate one standard deviation from the average

read pairs, independent of alignment approach, can

notably correct errors that accumulate in sequencing

instruments [55]

We also reported the application of MERIT to

ultra-deep sequencing data obtained from the amplification of

multiple clinically relevant loci using four high-fidelity

polymerase enzymes Although there is limited

varia-tion in both the rates of error and dependence on the

genomic content of the amplified region, our results

indicated that profiling polymerase misincorporation

pat-tern according to genomic context has important clinical

consequences Specifically, we showed that error rates

obtained from deep-sequencing of clinical specimens may

reflect processes that affect DNA quality during sample

preparations

Sample heterogeneity, especially when low-abundance mutations are present, can confound MERIT’s sequenc-ing error profiles Therefore, when MERIT is applied to clinical specimens from which true mutations are not removed, the estimated rates represent the upper bound

of true sequencing error rates Our results also demon-strated that assigning a single allele frequency threshold

to detect mutations may result in substantial false posi-tive as well as false negaposi-tive calls Not only were neigh-boring mutational hotspots in one gene affected with markedly different error rates, there was significant vari-ation in the sensitivity of detecting common amino acid changes within each residue These data suggested that some of these mutations may in fact be more prevalent

at sub-clonal levels in disease populations than previously